Email

cnievergelt@ucsd.edu

Suprateek Kundu, PhD


A. Research Question, Goal, or Specific Aims

This proposal develops state of the art approaches for addressing challenging questions related to the neurobiological mechanisms affecting clinical outcomes of interest in the presence of heterogeneity represented by underlying symptoms and other factors. We focus on developing integrative approaches for brain connectome-based analyses, which combines the multi-modal imaging (fMRI and diffusion MRI) of brain function and structure, clinical and behavioral measures, while accounting for heterogeneity across samples and incorporating prior biological knowledge in terms of the brain anatomy. Our goals involve important questions in neuroscience which have received limited or no attention so far, such as estimating dynamic brain connectivity while incorporating brain anatomical structure, and subsequently examining which dynamic functional connections drive the clinical outcome, accounting for heterogeneity in terms of disease sub-categories when predicting the clinical outcome based on brain measurements which lie on an underlying brain network, and investigating differences in shapes of white matter fiber bundles which drive the clinical outcome. To address such challenging goals, we develop state-of-the-art statistical approaches which incorporate significant innovations in terms of methodology and translational science.

Aim 1: To develop novel brain network modeling tools to investigate static and dynamic functional connectivity (FC) features that help distinguish heterogeneous PTSD subpopulations.

Aim 1.1: To examine stationary FC differences between PTSD patient sub-groups.

Aim 1.2. To discover dynamic brain functional neural circuits which regulate clinical outcomes.

Aim 2: Regression analysis incorporating brain network and accounting for PTSD subgroups. We develop an approach for jointly modeling multiple regression models (corresponding to different sub- population) for predicting the clinical outcome based on fMRI data and incorporating the subgroup- specific brain network characterizing the fMRI measurements.

Aim 3: Bayesian 3D-shape estimation of white matter fiber bundles and examining shape differences which drive the clinical outcome, which is expected to provide richer geometric information and more meaningful visualizations.

B. Analytic Plan

A number of brain regions have been implicated in PTSD and other anxiety disorders; however, none of these regions in isolation has been distinguished as the sole or discrete site responsible for the disorder pathology, and hence there is an increasing interest in the past decade to identify dysfunctional brain circuits which capture the underlying biological susceptibility towards PTSD symptoms. Our Aim 1 provides significant improvements over existing approaches which are

concerned with discovering brain network differences across cohorts with and without psychopathology and/or trauma exposure. Firstly in Aim 1.1, we examine the shared patterns and differences across brain networks corresponding to PTSD sub-groups comprising individuals with similar symptom scores. The subgroups potentially represent heterogeneity within each of the broader cohorts of subjects with and without PTSD and/or with and without trauma exposure. Comparing networks between such subgroups provides a more granular analysis and provides a fresh perspective regarding brain network differences driving PTSD heterogeneity, that may not be addressed by simply examining cohorts with and without psychopathology and/or trauma exposure, as is the case with standard approaches. Moreover, the networks across the subgroups are estimated jointly, resulting in greater power to detect shared patterns and differences [28] compared to standard approaches which separately estimate each network and subsequently compare these networks using mass-univariate hypothesis testing while controlling the family-wise error rate. These approaches often have reduced power to detect true differences due to the fact that they have to adjust for a massive number of multiple comparisons, and because they do not borrow information across networks, which results in less accurate estimates. The proposed method goes beyond the standard approaches for jointly estimating multiple networks that can not typically handle the scenario where it is of interest to jointly model multiple networks in terms of supplemental covariates (eg: symptom scores). Such network valued regression approaches has had limited progress in literature so far, with no applications to brain networks to our knowledge, making our contribution particularly significant.

Secondly in Aim 1.2, we investigate the role of dynamic brain networks in predicting continuous clinical outcomes of interest. An increasing number of recent findings have provided evidence on the dynamic nature of brain functional organization, and studies are beginning to identify potential correlates of temporal variations in connectivity in simultaneously recorded electrophysiological data, behavior, and disease status. Recent evidence has illustrated that dynamic networks are superior classifiers of PTSD compared to healthy controls. It is crucial that researchers start focusing on the assessment and analysis of dynamic, temporal data which may bring us closer to developing novel recommendations for intervention or prevention strategies for individuals focusing on personalized medicine. In one of the first of its kind methods, we propose a novel dynamic network modeling approach which incorporates the brain anatomical knowledge to accurately estimate the temporal dynamics for each subject, and provide a more precise characterization of short-lived brain states which are otherwise difficult to detect using a small number of time scans. In the second stage, we investigate the role of dynamic circuits (estimated in the first stage) in regulating the continuous clinical outcome of interest, via a scalar-on-function regression which also enables variable selection with the set of significant dynamic FCs potentially changing across subjects based on their symptom scores. In addition, the Both Aims 1.1 and 1.2 address clinically significant questions in literature which have received no or limited attention due to the lack of state-of-the-art statistical methodology needed to answer the scientific hypothesis of interest. Aim 2 makes significant contributions in developing a regression modeling approach for predicting a continuous clinical outcome of interest based on neuroimaging data, while accounting for heterogeneity (in terms of PTSD subgroups defined earlier), and also incorporating the brain network knowledge to inform the regression models. The proposed approach is able to model distinct relationships between the clinical outcome and neuroimaging biomarkers corresponding to the different PTSD subgroups, while pooling information across the subgroups and also being able to discover significant brain regions responsible for driving the

clinical out- come of interest in each subgroup. The latter is made possible by a variable selection approach which is informed by the subgroup-specific brain network information, and is estimated in Aim 1.1. From a method- ological perspective, our method is one of the first Bayesian approaches to jointly estimate multiple regression models involving high dimensional brain imaging measurements, while incorporating subgroup-specific bio- logical network knowledge. The joint analysis of different PTSD subgroups defined by symptom scales is meaningful since the symptoms are known to covary because they are linked via causal associations with each other, and incorporating prior network knowledge is expected to better account for the covariance structure in the spatially varying fMRI signals that will lead to more statistical power in detecting the true effects, which is especially important given the reduced sample size of certain PTSD subgroups.

Aim 3 addresses a key step in structural analysis of WM fiber tracts, and their use in predicting clinical observables, by proposing novel Bayesian approaches for accurate and efficient extraction of fiber bundles connecting diverse regions of interest. For this Aim, we will develop Bayesian shape-based priors to estimate white matter fiber bundles for each individual by defining appropriate energy functions, and then propose a shape- based regression analysis to discover differences in shapes of fiber bundles which influence the clinical outcome. If the data is not-preprocessed, standard pre-processing steps will be applied to fMRI and DTI data before analysis. Otherwise, pre-processed data will be used from ENIGMA consortium.The majority of studies reported a significant reduction in WM volume of major fiber tracts including the corpus callosum, the cingulum bundle as well as the left posterior cingulate, and changes in WM volume in adult-onset PTSD in comparison to healthy control subjects with or without traumatic experience was also discovered (see the recent reviews and meta-analysis in [69]-[70]). Changes in microstructural WM in individuals with adult-onset PTSD have been measured using manual tracing, volumetric morphometry and DTI, with DTI studies gaining an increasing popularity in recent years. These studies typically compare diffusivity measures (such as mean diffusivity, radial diffusivity, and functional anisotropy, and so on) between two or more populations at the voxel or ROI level and discover brain regions with differential diffusivity across different risk populations. These standard summary diffusivity measures only provides limited anatomical information and are not capable of capturing the richer geometric features of WM fiber tracts which are responsible for regulating the clinical outcome of interest. Furthermore, the existing analysis focus on comparing the brain anatomy be- tween pre-defined groups with categorical diagnosis of PTSD (eg: with and without psychopathology, or with and without trauma), which may be challenging given the difficulty of defining appropriate risk and control groups, and due to the underlying heterogeneity present in the PTSD population that may render the group level inferences inadequate. It is important to note that there are essentially no existing approaches to investigate the role of structural brain networks in the resilient functioning after trauma exposure, which is a clinical outcome of interest in PTSD. Unlike existing methods which focus on diffusivity measures or volume of WM, ours is one of the first studies to our knowledge which performs a shape-based analysis of 3-dimensional WM fiber bundles and their role in driving clinical outcomes. Such an analysis is expected to provide richer geometric information and will enable visualization of 3-dimensional fiber tracts responsible for heterogeneity in PTSD samples, but is involves several challenges. For example, we need to develop shape based anatomical priors to guide the estimation of 3-dimensional fiber tracts which is expected to lead to a smaller number of erroneous estimated tracts that is inconsistent with the underlying biology. The above features represent considerable and significant methodological and clinical contributions to the body of existing work and is expected to have a major impact.

C. Analytic Personnel

Dr. Suprateek Kundu (Associate Professor in Biostatistics at MD Anderson Cancer Research Center) will be responsible for leading the analysis. He will have expert help from Dr. Rajendra Morey at Duke (also a key member of the ENIGMA consortium), as well as co-investigators Dr. Jennifer Stevens and Dr. Negar Fani at Emory, who are experts in trauma exposure and PTSD symptoms as well as functional and structural imaging. Dr. Kundu will hire a research assistant to help with the analysis.

Recommended Articles

Leave A Comment

Your email address will not be published. Required fields are marked *